华为世界模型来了！单卡30分钟生成272㎡场景

Core Viewpoint - The article discusses the launch of WordGrow, a world model developed by Huawei in collaboration with Shanghai Jiao Tong University and Huazhong University of Science and Technology, capable of generating large indoor scenes with high realism and coherent geometry [1][2]. Group 1: Technology Overview - WordGrow can generate an indoor scene of 1800 square meters (19x39 blocks) in just 30 minutes on a single A100 GPU, achieving a speed six times faster than similar technologies [16][17]. - The model employs three core technologies: precise data preprocessing, a 3D block completion mechanism, and a coarse-to-fine generation strategy [10][12][14]. - The model's geometric reconstruction metrics, MMD and COV, have reached state-of-the-art levels, with a FID score as low as 7.52, significantly outperforming mainstream methods like SynCity and BlockFusion [17]. Group 2: Technical Details - The first step involves data preprocessing from large datasets like 3D-FRONT, ensuring high-quality sample extraction and scene segmentation [10]. - The second step focuses on seamless integration of 3D structures, maintaining consistent visual styles and eliminating issues like texture misalignment [12]. - The final step enhances scene resolution and detail by refining the overall layout and filling in missing elements such as furniture and textures [12][14]. Group 3: Performance Metrics - Experimental results indicate that even when expanded to a 7x7 block ultra-large scene, the edge quality remains stable [15]. - The model's performance metrics show a significant improvement over competitors, with MMD values of 0.97 and EMD values indicating superior quality [15][16]. Group 4: Team Background - The research was conducted by Sikuang Li and Chen Yang from Shanghai Jiao Tong University during their internship at Huawei, with guidance from renowned AI expert Tian Qi [18][19]. - Tian Qi is recognized as the Chief Scientist of Huawei's Terminal BG and an esteemed member of international scientific communities [20].